Experimental scheme intelligent recommendation method and system for solving cold start problem

Abstract

This invention discloses an intelligent recommendation method and system for experimental schemes to solve the cold start problem, belonging to the field of experimental recommendation technology. The method includes: acquiring limited interaction data of the target cold-start user and the physicochemical properties and molecular structure characteristics of the reagents to be recommended; calculating the similarity with known mature experimental schemes based on the properties and characteristics to construct an initial user preference model; recommending probe experiments using an active learning strategy and obtaining user feedback; based on a meta-learning framework, treating each cold-start user as an independent task, rapidly updating parameters using feedback data and optimizing initialization parameters at the cross-user level to obtain a personalized recommendation model; and outputting an executable script adapted to laboratory automated robots. This invention, by integrating physicochemical property analysis, active learning, and meta-learning mechanisms, can effectively solve the cold start problem in experimental recommendation, achieving accurate personalized recommendations and seamless integration with automated equipment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of experimental recommendation, in particular to an experimental scheme intelligent recommendation method and system for solving the cold start problem. BACKGROUND

[0002] With the rapid development of laboratory automation and artificial intelligence technology, intelligent experimental scheme recommendation systems play an increasingly important role in scientific research and production. However, traditional recommendation systems generally have a serious cold start problem when facing new experiments or new materials. Due to the lack of historical experimental data, collaborative filtering-based methods are difficult to accurately model user preferences, resulting in poor recommendation results for new users. In particular, in the field of experimental science, the physical and chemical properties of experimental reagents are complex and diverse, and user preferences have high professional and personalized characteristics, making it difficult for traditional recommendation methods to extract effective information from limited interactions.

[0003] To solve the above problems, researchers have proposed cold start recommendation methods based on meta-learning. For example, Chinese patent application CN113836393A discloses a cold start recommendation method based on preference adaptive meta-learning, which represents user preferences by constructing a heterogeneous information network and an attention mechanism. However, existing methods still have the following shortcomings: first, they ignore the relationship between users, resulting in poor model generalization ability; second, they lack an effective mechanism to discover individual preferences from limited interactions of cold start users; third, they fail to fully utilize the physical and chemical properties and molecular structure characteristics of experimental reagents, and their performance significantly decreases in scenarios where auxiliary information is not available. In addition, existing technologies only stop at the level of scheme recommendation and fail to achieve seamless integration with laboratory automation equipment, which needs to be improved in practicality. SUMMARY

[0004] To solve the above technical problems in related technologies, the present application proposes an experimental scheme intelligent recommendation method and system for solving the cold start problem, which can overcome the above shortcomings of existing technologies.

[0005] To achieve the above technical purposes, the technical solution of the present application is as follows: An experimental scheme intelligent recommendation method for solving the cold start problem; The experimental scheme intelligent recommendation method for solving the cold start problem comprises the following steps: Obtain limited interaction data of a target cold start user and physical and chemical properties and molecular structure characteristics of experimental reagents to be recommended; Based on the physical and chemical properties and molecular structure characteristics, calculate the similarity between the experimental schemes to be recommended and known mature experimental schemes, and construct an initial user preference model; Use an active learning strategy to recommend a preset number of probe experiments to the target cold start user, and obtain feedback data of the user on the probe experiments; Based on the meta-learning framework, each cold-start user is treated as an independent task. The feedback data is used as a support set for rapid parameter updates, and the initialization parameters are optimized at the cross-user level to obtain a personalized recommendation model. The personalized recommendation model outputs an executable script adapted to laboratory automation robots.

[0006] Furthermore, the molecular structure features are obtained through molecular fingerprint encoding or graph neural network modeling; the construction of the initial user preference model further includes: embedding the user's basic attribute features into a representation, and using an attention mechanism to assign weights to the heterogeneous attribute features of the items, and weighted summing to obtain the initial vector representation of the items; wherein, when using graph neural network modeling, the molecular formula of the experimental reagent is converted into a molecular graph structure, with atoms as nodes and chemical bonds as edges, and the topological features of the molecular graph are extracted through a graph convolutional network or a graph attention network to generate a molecular structure latent vector.

[0007] Furthermore, the process of optimizing the initialization parameters based on the meta-learning framework includes: Define the support set as the feedback data of cold-start users on the probe experiment, and the query set as the potential preferences of users for other experimental schemes; In the inner loop, for a specific cold-start user or experiment, the local gradient is quickly calculated and the parameters are updated using the support set. In the outer loop, the initialization parameters are optimized across multiple user tasks, enabling the model to reach an optimized state with a small number of iterations when faced with new experiments. Meanwhile, a cross-user meta-knowledge transfer mechanism is introduced to map the preference patterns of active users to the prior parameters of cold-start users through the meta-network.

[0008] Furthermore, the personalized recommendation model is constructed based on an improved dual-tower structure, including a user tower and a material tower; wherein, the input of the user tower includes basic user attribute features and user embedding information, and the input of the material tower includes the physicochemical properties and molecular structure features of experimental reagents; the outputs of the two towers are connected and then jointly trained. The training process of the dual-tower model includes: inputting user interaction data sequentially into the embedding layer, user ID vector generator, user tower and material tower to obtain a first benchmark model; obtaining a second cold start recall model based on the first benchmark model, and updating the second cold start recall model through update parameters issued by the server to obtain the first cold start recall model.

[0009] Furthermore, the executable script includes experimental steps, experimental parameters, and experimental time information, which are used to directly control the laboratory automation equipment to execute the recommended scheme.

[0010] According to another aspect of the present invention, an intelligent recommendation system for experimental schemes to solve the cold start problem is provided; The intelligent recommendation system for solving the cold start problem includes: The data acquisition module is used to acquire limited interaction data of the target cold-start user and the physicochemical properties and molecular structure characteristics of the experimental reagents to be recommended; The initial preference modeling module is used to calculate the similarity between the experimental scheme to be recommended and the known mature experimental schemes based on the physicochemical properties and molecular structure characteristics, and to construct an initial user preference model. The initial preference modeling module includes a feature extraction unit, which is used to model molecular structural features through molecular fingerprint encoding or graph neural networks; An active learning module is used to recommend a preset number of probe experiments to the target cold-start user and obtain user feedback data on the probe experiments; The meta-learning training module is used to treat each cold-start user as an independent task based on the meta-learning framework, use the feedback data as a support set to quickly update the parameters, and optimize the initialization parameters at the cross-user level to obtain a personalized recommendation model. The script output module is used to output an executable script adapted to the laboratory automation robot based on the personalized recommendation model.

[0011] Furthermore, the meta-learning training module includes: Support set and query set building units are used to define the support set as the feedback data of cold-start users to the probe experiment, and the query set as the user's potential preferences for other experimental schemes; The inner loop update unit is used to quickly calculate local gradients and update parameters using the support set for specific cold-start users or experiments. The outer loop optimization unit is used to optimize the initialization parameters across multiple user tasks, enabling the model to reach an optimized state with a small number of iterations when faced with new experiments. The cross-user knowledge transfer unit is used to map the preference patterns of active users to the prior parameters of cold-start users through the meta-network.

[0012] Furthermore, when the feature extraction unit uses graph neural network modeling, it converts the molecular formula of the experimental reagent into a molecular graph structure, extracts the topological features of the molecular graph with atoms as nodes and chemical bonds as edges, and generates a latent vector of the molecular structure.

[0013] Furthermore, the personalized recommendation model is constructed based on an improved dual-tower structure, including a user tower and a material tower; wherein, the input of the user tower includes basic user attribute features and user embedding information, and the input of the material tower includes the physicochemical properties and molecular structure features of experimental reagents; the outputs of the two towers are connected and then jointly trained. The training process of the dual-tower model includes: inputting user interaction data sequentially into the embedding layer, user ID vector generator, user tower and material tower to obtain a first benchmark model; obtaining a second cold start recall model based on the first benchmark model, and updating the second cold start recall model through update parameters issued by the server to obtain the first cold start recall model.

[0014] Furthermore, the executable script includes experimental steps, experimental parameters, and experimental time information, which are used to directly control the laboratory automation equipment to execute the recommended scheme.

[0015] The beneficial effects of this invention are as follows: By analyzing the similarity between the physicochemical properties of experimental reagents and the molecular structure characteristics, the cold start problem when there is no historical data for new experiments is effectively solved, thereby improving the accuracy and reliability of recommendations; by adopting an active learning strategy to obtain user feedback on the detection experiments, the model parameters can be quickly corrected, thereby improving the model's adaptability and generalization ability; by realizing cross-user knowledge transfer based on a meta-learning framework, user preferences can be mined from limited interactions, thereby achieving the effect of accurate personalized recommendations; by outputting execution scripts adapted to laboratory automated robots, seamless connection from experimental plans to actual operations can be achieved, thereby improving the practicality and convenience of recommendations. Detailed Implementation

[0016] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0017] The intelligent recommendation method for solving the cold start problem according to an embodiment of the present invention includes the following steps: Acquire limited interaction data of the target cold-start users and the physicochemical properties and molecular structure characteristics of the experimental reagents to be recommended; Based on the aforementioned physicochemical properties and molecular structure characteristics, the similarity between the experimental scheme to be recommended and the known mature experimental schemes is calculated, and an initial user preference model is constructed. An active learning strategy is adopted to recommend a preset number of probing experiments to target cold-start users and obtain user feedback data on the probing experiments. Based on the meta-learning framework, each cold-start user is treated as an independent task. The feedback data is used as a support set for rapid parameter updates, and the initialization parameters are optimized at the cross-user level to obtain a personalized recommendation model. The personalized recommendation model outputs an executable script adapted to laboratory automation robots.

[0018] According to an embodiment of the present invention, the intelligent recommendation method for experimental schemes to solve the cold start problem, in a specific implementation, the molecular structure features are obtained through molecular fingerprint encoding or graph neural network modeling; the construction of the initial user preference model further includes: embedding the user's basic attribute features, and using an attention mechanism to assign weights to the heterogeneous attribute features of items, and weighted summing to obtain the initial vector representation of the items; wherein, when using graph neural network modeling, the molecular formula of the experimental reagent is converted into a molecular graph structure, with atoms as nodes and chemical bonds as edges, and the topological features of the molecular graph are extracted through a graph convolutional network or a graph attention network to generate a latent vector of molecular structure.

[0019] According to an embodiment of the present invention, the intelligent recommendation method for experimental schemes to solve the cold start problem, in a specific implementation, includes the process of optimizing initialization parameters based on a meta-learning framework as follows: Define the support set as the feedback data of cold-start users on the probe experiment, and the query set as the potential preferences of users for other experimental schemes; In the inner loop, for a specific cold-start user or experiment, the local gradient is quickly calculated and the parameters are updated using the support set. In the outer loop, the initialization parameters are optimized across multiple user tasks, enabling the model to reach an optimized state with a small number of iterations when faced with new experiments. Meanwhile, a cross-user meta-knowledge transfer mechanism is introduced to map the preference patterns of active users to the prior parameters of cold-start users through the meta-network.

[0020] According to an embodiment of the present invention, the intelligent recommendation method for solving the cold start problem is described in a specific implementation whereby the personalized recommendation model is constructed based on an improved dual-tower structure, including a user tower and a material tower. The user tower's input includes basic user attribute features and user embedding information, while the material tower's input includes the physicochemical properties and molecular structure features of the experimental reagents. The outputs of the two towers are connected and then jointly trained. The training process of the dual-tower model includes: inputting user interaction data sequentially into the embedding layer, user ID vector generator, user tower and material tower to obtain a first benchmark model; obtaining a second cold start recall model based on the first benchmark model, and updating the second cold start recall model through update parameters issued by the server to obtain the first cold start recall model.

[0021] According to an embodiment of the present invention, the intelligent recommendation method for solving the cold start problem includes, in a specific implementation, an executable script comprising experimental steps, experimental parameters, and experimental time information, which is used to directly control laboratory automation equipment to execute the recommended scheme.

[0022] Secondly, the intelligent recommendation system for experimental schemes to solve the cold start problem according to embodiments of the present invention includes: The data acquisition module is used to acquire limited interaction data of the target cold-start user and the physicochemical properties and molecular structure characteristics of the experimental reagents to be recommended; The initial preference modeling module is used to calculate the similarity between the experimental scheme to be recommended and the known mature experimental schemes based on the physicochemical properties and molecular structure characteristics, and to construct an initial user preference model. The initial preference modeling module includes a feature extraction unit, which is used to model molecular structural features through molecular fingerprint encoding or graph neural networks; An active learning module is used to recommend a preset number of probe experiments to the target cold-start user and obtain user feedback data on the probe experiments; The meta-learning training module is used to treat each cold-start user as an independent task based on the meta-learning framework, use the feedback data as a support set to quickly update the parameters, and optimize the initialization parameters at the cross-user level to obtain a personalized recommendation model. The script output module is used to output an executable script adapted to the laboratory automation robot based on the personalized recommendation model.

[0023] In a specific embodiment of the intelligent recommendation system for experimental schemes to solve the cold start problem according to an embodiment of the present invention, the meta-learning training module includes: Support set and query set building units are used to define the support set as the feedback data of cold-start users to the probe experiment, and the query set as the user's potential preferences for other experimental schemes; The inner loop update unit is used to quickly calculate local gradients and update parameters using the support set for specific cold-start users or experiments. The outer loop optimization unit is used to optimize the initialization parameters across multiple user tasks, enabling the model to reach an optimized state with a small number of iterations when faced with new experiments. The cross-user knowledge transfer unit is used to map the preference patterns of active users to the prior parameters of cold-start users through the meta-network.

[0024] In a specific embodiment of the intelligent recommendation system for solving the cold start problem according to an embodiment of the present invention, when the feature extraction unit uses graph neural network modeling, it converts the molecular formula of the experimental reagent into a molecular graph structure, extracts the topological features of the molecular graph with atoms as nodes and chemical bonds as edges, and generates a latent vector of the molecular structure.

[0025] According to an embodiment of the present invention, the intelligent recommendation system for solving the cold start problem of the experimental scheme, in a specific implementation, the personalized recommendation model is constructed based on an improved dual-tower structure, including a user tower and a material tower; wherein, the input of the user tower includes the user's basic attribute features and user embedding information, and the input of the material tower includes the physicochemical properties and molecular structure features of the experimental reagents; the outputs of the two towers are connected and then jointly trained. The training process of the dual-tower model includes: inputting user interaction data sequentially into the embedding layer, user ID vector generator, user tower and material tower to obtain a first benchmark model; obtaining a second cold start recall model based on the first benchmark model, and updating the second cold start recall model through update parameters issued by the server to obtain the first cold start recall model.

[0026] According to an embodiment of the present invention, the intelligent recommendation system for solving the cold start problem includes, in a specific implementation, an executable script comprising experimental steps, experimental parameters, and experimental time information, which is used to directly control laboratory automation equipment to execute the recommended scheme.

[0027] To facilitate understanding of the above technical solutions of the present invention, the following embodiments will be used to describe the above technical solutions of the present invention in detail.

[0028] Example 1 This embodiment provides an intelligent recommendation method for experimental schemes to solve the cold start problem. The specific implementation steps are as follows: Step 1: Obtain the experimental dataset. Due to the lack of historical experimental data, publicly available experimental datasets, such as DoubanBook and MovieLens, can be used in the initial stage of the system. The experimental dataset contains user information, item information, and user ratings for items. For experimental datasets lacking rating information, randomly generated rating data can be used.

[0029] Step 2: Data Preprocessing. The acquired experimental dataset is randomly partitioned, for example, into an 8:2 training and test set. The training set is used to train the meta-learning model, and the test set is used to evaluate model performance. Further, the training set is divided into four categories: existing users-existing items, cold-start users-existing items, cold-start items-existing users, and cold-start users-cold-start items. Existing users refer to users who have used the application; cold-start users refer to newly registered users; existing items refer to items purchased by users within the application; and cold-start items refer to newly released items.

[0030] Step 3: Feature Extraction and Processing. First, feature extraction is performed on users and items. For users, basic attribute features such as age, gender, and occupation, as well as user embedding information, are extracted. For items, physicochemical properties and molecular structure features are extracted. Then, the features are normalized to ensure the feature values ​​are between 0 and 1.

[0031] Step 4: Establish the initial vector representation for the cold-start user. Based on the user's basic attribute features, a vanilla attention mechanism is used to assign different weights to the item's attribute features, and the weighted sums are used to obtain the item's initial vector representation. Specifically, for the user, the corresponding feature embedding matrix is ​​used to extract the user's attribute feature set, and the vector representations corresponding to each attribute feature are concatenated to obtain the user's initial vector representation. For items, since item attributes are usually heterogeneous, a vanilla attention mechanism is used to assign different weights to different attribute features, and the weighted sums are used to obtain the vector representations of the attribute features.

[0032] Step 5: Establish a preference model for cold-start users. Employing an active learning strategy, the system proactively recommends several key "probe experiments." Through feedback from a small number of real experiments, the recommendation algorithm parameters are quickly adjusted. Specifically, user and item data from the training set are sequentially input into the embedding layer, user tower, and item tower of the improved dual-tower model for training, resulting in a first baseline model. Then, a second cold-start recall model is obtained based on the first baseline model. This second cold-start recall model is updated using update parameters issued by the server, resulting in a first cold-start recall model. Finally, the first cold-start recall model is used to process target cold-start users, yielding a first vector.

[0033] Step 6: Build a personalized model for cold-start users. The constructed user dataset is sequentially input into the embedding layer, user ID vector generator, user tower, and material tower of the improved dual-tower model for training, resulting in the first baseline model. By utilizing the user ID vector generator to leverage users' basic attributes, the expressive power of the second cold-start recall model is improved, enabling the model to simultaneously capture the interests of both active and cold-start users.

[0034] Step 7: Output Recommendation Results. Based on the established personalized model of the cold-start user, output an executable script adapted to the laboratory automated robot. Specifically, for each recommended experimental plan, the system generates a corresponding execution script, achieving a seamless connection from experimental plan to actual operation.

[0035] Example 2 Based on Example 1, this example further provides specific implementation steps for feature extraction and processing: Step 31: Feature Extraction. First, feature extraction is performed on the user, extracting basic attribute features such as age, gender, and occupation. One-hot encoding is used to represent these basic user attributes. For example, gender can be represented as a 0 / 1 vector, and age as a 10 / 20 / 30 vector. Next, feature extraction is performed on the items, extracting physicochemical properties and molecular structure features. For physicochemical properties, features such as pH value, solubility, and stability are extracted. For molecular structure features, molecular fingerprinting or graph neural networks can be used for modeling. When using graph neural networks, the molecular formula of each experimental reagent is converted into a molecular graph structure, where atoms are nodes and chemical bonds are edges. The topological features of the molecular graph are extracted using a graph convolutional network or graph attention network to generate a latent molecular structure vector. This latent vector not only serves as input to the dual-tower model for item features but also as prior knowledge in the meta-learning initialization phase, used for generating initial vectors for new experimental reagents in a cold-start state.

[0036] Step 32: Feature Normalization. The extracted features are normalized so that the feature values ​​are between 0 and 1. Specifically, for user features, the Min-Max normalization method is used. For item features, the Z-Score normalization method is used.

[0037] Step 33: Establish the initial vector representation of the cold-start user. Based on the user's basic attribute features, a vanilla attention mechanism is used to assign different weights to the attribute features of the item, and the weighted sum is used to obtain the initial vector representation of the item. Specifically, for the user, the set of attribute features of the user is extracted using the corresponding feature embedding matrix, and the vector representations corresponding to each attribute feature are concatenated to obtain the user's initial vector representation. For the item, since the attributes of the item are usually heterogeneous, a vanilla attention mechanism is used to assign different weights to different attribute features, and the weighted sum is used to obtain the vector representation of the attribute features. For example, for an item, its molecular structure features include sour taste, polarity, solubility, etc., and attention weights of 0.2, 0.3, and 0.5 are used respectively.

[0038] Step 34: Establish a preference model for cold-start users. Employing an active learning strategy, the system proactively recommends several key "probe experiments." Through feedback from a small number of real experiments, the recommendation algorithm parameters are quickly adjusted. Specifically, user and item data from the training set are sequentially input into the embedding layer, user tower, and item tower of the improved dual-tower model for training, resulting in a first baseline model. Then, a second cold-start recall model is obtained based on the first baseline model. This second cold-start recall model is updated using update parameters sent from the server, resulting in a first cold-start recall model. This first cold-start recall model is then used to process target cold-start users, yielding a first vector. For example, the system proactively recommends 10 probe experiments for cold-start users to evaluate. After users provide feedback and ratings, the system automatically adjusts the weight parameters in the recommendation algorithm based on the rating differences.

[0039] Example 3 Based on Example 1, this example further provides the specific implementation steps of the cold start user personalization model: Step 61: Build the user embedding model. The constructed user dataset is sequentially input into the embedding layer, user ID vector generator, user tower, and material tower of the improved dual-tower model for training, resulting in the first baseline model. Specifically, a deep neural network is used as the embedding layer, with the number of input layer nodes equal to the user feature dimension, 64 hidden layer nodes, and 128 output layer nodes. The user ID vector generator uses a recurrent neural network (RNN) structure, with 256 hidden layer nodes and 128 output layer nodes.

[0040] Step 62: Establish the material embedding model. For items, another dual-tower model is used for training. The input of this dual-tower model is the item features, the number of hidden layer nodes is 64, and the number of output layer nodes is 128. Note that this material embedding model is independent of the previous user embedding model.

[0041] Step 63: Establish a personalized model for the cold-start user. Connect the outputs of the two embedding models to form a dual-tower model. The user tower's input is the user embedding, and its output is the user latent vector; the material tower's input is the item embedding, and its output is the item latent vector. The connection between the two towers is achieved through concatenation. The entire model adopts a Siamese network structure, sharing the parameters of both the user tower and the material tower.

[0042] Step 64: Train the personalized model for cold-start users. The Adam optimization algorithm is used with a learning rate of 0.001, a batch size of 64, and 1000 iterations. In each iteration, 32 training samples are randomly selected for training, and the remaining samples are used as a cache pool for subsequent adaptive learning.

[0043] Step 65: Adaptive Learning. Every 10 epochs, the model undergoes adaptive learning using samples from the cache pool. Specifically, the samples in the cache pool are randomly shuffled, and the first 32 samples are used as input for adaptive learning to fine-tune the model. The learning rate for adaptive learning is 0.0001, and only the classification layer is updated.

[0044] Step 66: Output Recommendation Results. Based on the established personalized model of the cold-start user, output an executable script adapted to the laboratory automation robot. Specifically, for each recommended experimental plan, the system generates a corresponding execution script, achieving seamless integration from experimental plan to actual operation. For example, for a recommended experimental plan, the system-generated execution script includes detailed information such as experimental steps, experimental parameters, and experimental time, directly applicable to the execution of laboratory automation equipment.

[0045] In summary, by utilizing the technical solutions described above, the similarity between the physicochemical properties of experimental reagents and the calculation schemes for molecular structural features is analyzed, effectively solving the cold start problem when new experiments lack historical data, thereby improving the accuracy and reliability of recommendations. An active learning strategy is employed to obtain user feedback on the detection experiments, thereby quickly correcting model parameters and enhancing the model's adaptability and generalization ability. A meta-learning framework enables cross-user knowledge transfer, thereby mining user preferences from limited interactions and achieving precise personalized recommendations. Furthermore, by outputting execution scripts adapted to laboratory automation robots, seamless integration from experimental plans to actual operations is achieved, further improving the practicality and convenience of recommendations.

[0046] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent recommendation method for experimental schemes to solve the cold start problem, characterized in that, Includes the following steps: Acquire limited interaction data of the target cold-start users and the physicochemical properties and molecular structure characteristics of the experimental reagents to be recommended; Based on the aforementioned physicochemical properties and molecular structure characteristics, the similarity between the experimental scheme to be recommended and the known mature experimental schemes is calculated, and an initial user preference model is constructed. An active learning strategy is adopted to recommend a preset number of probing experiments to target cold-start users and obtain user feedback data on the probing experiments. Based on the meta-learning framework, each cold-start user is treated as an independent task. The feedback data is used as a support set for rapid parameter updates, and the initialization parameters are optimized at the cross-user level to obtain a personalized recommendation model. The personalized recommendation model outputs an executable script adapted to laboratory automation robots.

2. The intelligent recommendation method for experimental schemes to solve the cold start problem according to claim 1, characterized in that, The molecular structure features are obtained through molecular fingerprint encoding or graph neural network modeling; the construction of the initial user preference model further includes: embedding the user's basic attribute features into a representation, and using an attention mechanism to assign weights to the heterogeneous attribute features of the items, and weighted summing to obtain the initial vector representation of the items; wherein, when using graph neural network modeling, the molecular formula of the experimental reagent is converted into a molecular graph structure, with atoms as nodes and chemical bonds as edges, and the topological features of the molecular graph are extracted through a graph convolutional network or a graph attention network to generate a latent vector of the molecular structure.

3. The intelligent recommendation method for experimental schemes to solve the cold start problem according to claim 1, characterized in that, The process of optimizing initialization parameters based on the meta-learning framework includes: Define the support set as the feedback data of cold-start users on the probe experiment, and the query set as the potential preferences of users for other experimental schemes; In the inner loop, for a specific cold-start user or experiment, the local gradient is quickly calculated and the parameters are updated using the support set. In the outer loop, the initialization parameters are optimized across multiple user tasks, enabling the model to reach an optimized state with a small number of iterations when faced with new experiments. Meanwhile, a cross-user meta-knowledge transfer mechanism is introduced to map the preference patterns of active users to the prior parameters of cold-start users through the meta-network.

4. The intelligent recommendation method for experimental schemes to solve the cold start problem according to claim 1, characterized in that, The personalized recommendation model is based on an improved dual-tower structure, including a user tower and a material tower. The user tower's input includes basic user attribute features and user embedding information, while the material tower's input includes the physicochemical properties and molecular structure features of experimental reagents. The outputs of the two towers are connected and then jointly trained. The training process of the dual-tower model includes: inputting user interaction data sequentially into the embedding layer, user ID vector generator, user tower and material tower to obtain a first benchmark model; obtaining a second cold start recall model based on the first benchmark model, and updating the second cold start recall model through update parameters issued by the server to obtain the first cold start recall model.

5. The intelligent recommendation method for experimental schemes to solve the cold start problem according to claim 1, characterized in that, The executable script includes experimental steps, experimental parameters, and experimental time information, and is used to directly control laboratory automation equipment to execute the recommended scheme.

6. An intelligent recommendation system for experimental schemes to solve the cold start problem, characterized in that: include: The data acquisition module is used to acquire limited interaction data of the target cold-start user and the physicochemical properties and molecular structure characteristics of the experimental reagents to be recommended; The initial preference modeling module is used to calculate the similarity between the experimental scheme to be recommended and the known mature experimental schemes based on the physicochemical properties and molecular structure characteristics, and to construct an initial user preference model. The initial preference modeling module includes a feature extraction unit, which is used to model molecular structural features through molecular fingerprint encoding or graph neural networks; An active learning module is used to recommend a preset number of probe experiments to the target cold-start user and obtain user feedback data on the probe experiments; The meta-learning training module is used to treat each cold-start user as an independent task based on the meta-learning framework, use the feedback data as a support set to quickly update the parameters, and optimize the initialization parameters at the cross-user level to obtain a personalized recommendation model. The script output module is used to output an executable script adapted to the laboratory automation robot based on the personalized recommendation model.

7. The intelligent recommendation system for experimental schemes to solve the cold start problem according to claim 6, characterized in that, The meta-learning training module includes: Support set and query set building units are used to define the support set as the feedback data of cold-start users to the probe experiment, and the query set as the user's potential preferences for other experimental schemes; The inner loop update unit is used to quickly calculate local gradients and update parameters using the support set for specific cold-start users or experiments. The outer loop optimization unit is used to optimize the initialization parameters across multiple user tasks, enabling the model to reach an optimized state with a small number of iterations when faced with new experiments. The cross-user knowledge transfer unit is used to map the preference patterns of active users to the prior parameters of cold-start users through the meta-network.

8. The intelligent recommendation system for experimental schemes to solve the cold start problem according to claim 6, characterized in that, When the feature extraction unit uses graph neural network modeling, it converts the molecular formula of the experimental reagent into a molecular graph structure, uses atoms as nodes and chemical bonds as edges, extracts the topological features of the molecular graph, and generates a latent vector of the molecular structure.

9. The intelligent recommendation system for experimental schemes to solve the cold start problem according to claim 6, characterized in that, The personalized recommendation model is based on an improved dual-tower structure, including a user tower and a material tower. The user tower's input includes basic user attribute features and user embedding information, while the material tower's input includes the physicochemical properties and molecular structure features of experimental reagents. The outputs of the two towers are connected and then jointly trained. The training process of the dual-tower model includes: inputting user interaction data sequentially into the embedding layer, user ID vector generator, user tower and material tower to obtain a first benchmark model; obtaining a second cold start recall model based on the first benchmark model, and updating the second cold start recall model through update parameters issued by the server to obtain the first cold start recall model.

10. The intelligent recommendation system for experimental schemes to solve the cold start problem according to claim 6, characterized in that, The executable script includes experimental steps, experimental parameters, and experimental time information, and is used to directly control laboratory automation equipment to execute the recommended scheme.